ID=151

• Dataset: 6n4o
• Software: PMEMD.cuda.MPI (amber/20.9-20.15-gomklc-2020a-avx512)
• Resource: 3 tasks, 1 cores, nodes, 3 GPUs, no NVLink
• CPU: Xeon E5-2650 (Sandy Bridge), 2.2 GHz
• GPU: Tesla-P100-PCIE-12GB, 16 cores/GPU
• Simulation speed: 7.97 ns/day
• Efficiency: 17.8 %
• Site: Cedar
• Date: Feb. 7, 2022, 11:35 p.m.
• Submission script:

  #!/bin/bash
  #SBATCH --nodes=1 --ntasks=3 --gpus-per-node=p100:3
  #SBATCH --mem-per-cpu=2000 --time=1:0:0
  
  INPFILE=pmemd_prod.in
  STEPS=40000
  # End of user input
  LOGFILE=production_${SLURM_JOBID}.log 
  module --force purge
  ml StdEnv/2020  gcc/9.3.0 cuda/11.0 openmpi/4.0.3 amber/20.9-20.15
  # Print resource info
  echo ${SLURM_NODELIST} running on ${SLURM_NTASKS} tasks
  cat /proc/cpuinfo | grep "model name" | uniq
  # Run simulation 
  srun pmemd.cuda.MPI -O -i $INPFILE -o $LOGFILE -p prmtop.parm7 -c restart.rst7
  #rm  $TMPFILE $LOGFILE mdinfo mdcrd restrt

• Notes:

  Inconsistent performance. One day on nodes, cdr340 and cdr341 speed was only about 9 ns/day. Two days later speed was back to normal.

  Compare jobs 22449024 and 22503974. Note abnormal RAM usage of the slow job 22449024. Nvidia-smi showed only three running processes for 22449024, while normal jobs had four.

  Script to calculate speed:

  #!/bin/bash
  STEPS=40000
  grep "NonSetup CPU time" production_*.log > tmp
  awk -v steps=$STEPS '{ print 3.6*2.4*steps*0.01/$6 }' tmp
  rm tmp

• Simulation input file:

  Benchmark
   &cntrl
    imin=0,irest=1,ntx=5,
    nstlim=20000,dt=0.001,
    ntc=2,ntf=2,
    cut=8.0, 
    ntpr=100, ntwx=1000,
    ntb=2, ntp=1, taup=2.0,
    ntt=3, gamma_ln=2.0, temp0=300.0,
   /
   &ewald
    nfft1=128,
    nfft2=128,
    nfft3=128,
   /